Conformational autoencoder for heterochiral peptide macrocycles, sampling low-energy 6- to 9-residue mainchain scaffolds for therapeutic design.
No providers recorded yet. Browse all providers
Peptide macrocycles containing D-amino acids and other non-canonical building blocks reach folds no natural protein does, which is why they are attractive therapeutics — and why models trained on the Protein Data Bank cannot design them. AfCycDesign, RFpeptides and HighFold2 inherit their training distribution from PDB proteins built of L-amino acids, so they work best on L-homochiral macrocycles. For heterochiral chemistry the state of the art remains physics-based sampling: Rosetta's Generalized Kinematic Closure (GeneralizedKIC), which solves analytically for the internal degrees of freedom needed to close a ring. Mapping an 8-residue macrocycle's landscape that way can take 50,000 samples, repeated again to validate every designed sequence.
CyclicCAE, from the Flatiron Institute's Center for Computational Biology and the University of Oregon, takes a different view of what the network is for. Rather than learning biology from experimental data, it is a fast function approximator for a costly simulation: the autoencoder is trained once on a GeneralizedKIC-generated ensemble, after which sampling that learned distribution of closed, low-energy conformations is cheap and reusable across many design campaigns. The encoder maps a vector of mainchain dihedrals, bond angles and bond lengths into a six-dimensional latent space; one decoder limb reconstructs geometry, the other regresses conformational energy.
That second limb makes the latent space a search space. Because energy is differentiable with respect to the latent coordinate, a conformation is relaxed by gradient descent along the manifold of accessible geometries rather than in Cartesian space, and extra loss terms — terminal bond geometry, a motif, a repulsion against a target's pocket atoms — are added at inference without retraining.
--l-homochiral flag when only L backbones are wanted.CyclicCAE is a 1D convolutional autoencoder: convolutional and MLP blocks with positional
encodings map the mainchain degree-of-freedom vector to a six-dimensional latent space in
every trained instance, and separate MLP and deconvolutional limbs emit predicted energies and
reconstructed geometry. One instance per ring size was trained on synthetic ensembles from
Rosetta's simple_cycpep_predict — 50,000 poly-glycine conformations
each for 6-, 7- and 8-mers and 100,000 for 9-mers — geometry-optimized with GAMESS and scored
with both Rosetta's ref2015 force field and the DFTB semi-empirical method.
Encoding and decoding 37 PDB macrocycle structures, none in training, gave median mainchain heavy-atom RMSDs below 1 Å for 6-, 7- and 8-mers and 1.8 Å for 9-mers; ten steps of latent-space gradient descent cut the worst 9-mer outlier from 2.587 to 1.935 Å. Scaffolds passed through Rosetta FastDesign yielded stably folding sequences (Pnear ≥ 0.8) far more often than GeneralizedKIC scaffolds for 7-, 8- and 9-mers, a one to two order of magnitude reduction in sampling time per usable design; raw sampling ran 2.2–3.5× faster on one CPU core and 5.7–8.2× faster on an A100 GPU. The pocket term cut clashes from 4% to 0% for NDM1 and from 25% to 6% for the C5AR1 GPCR at a 0.5 Å threshold.
The model targets macrocyclic peptide drug discovery, where the bottleneck is proposing closed backbones worth designing sequences onto. It supplies scaffolds for Rosetta design runs, relaxes existing structures, diversifies a promising conformation by MCMC, rigidifies a linear binding motif inside a small macrocycle, and proposes backbones shaped to a pocket such as NDM1 or a GPCR — work that previously took hundreds of parallel GeneralizedKIC runs on a cluster.
CyclicCAE shows that deep learning can help in a domain with almost no experimental training data, by learning the output distribution of an expensive physics-based sampler instead of learning from nature. The authors note the architecture is specific neither to these ring sizes, nor to poly-glycine, nor even to α-amino acids, so peptoid macrocycles are reachable with suitable training data. The released instances cover only 6- to 9-residue N-to-C amide-cyclized rings, and the model is sequence-unaware, so it misses conformations stabilized by bulky non-canonical sidechains. Architecture, training scripts and weights carry a BSD 3-clause licence, and the work is a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.